The Use of Color in Multidimensional Graphical Information Display

Transcription

1 The Use of Color in Multidimensional Graphical Information Display Ethan D. Montag Munsell Color Science Loratory Chester F. Carlson Center for Imaging Science Rochester Institute of Technology, Rochester, NY Abstract The performance in judging values in a univariate map encoded using five different color scales was tested in eleven subjects. Digital elevation maps (DEMs) were encoded using: ) an RGB gray scale (RGB), 2) a gray scale based on CIELAB (), 3) a scale with an added red hue component (Red ), 4) an scale with continuous hue change (Spectral ), and 5) a gray scale based on luminance (Luminance). Performance was tested using an Evaluation task and a Production task. For both tasks judgments were made both with and without legends for all five encoding schemes. The results show a significant effect of choice of encoding scheme, the presence or sence of a legend, and an interaction between these two factors. Performance with a legend was significantly better than without one. The Spectral scale led to the best performance while Luminance encoding was the worst. This experiment is a first step in using quantifile psychophysical procedures to evaluate the effectiveness of different color encoding schemes on the interpretility of multidimensional graphical images. Keywords: visualization, color imaging, psychophysics Introduction The last step, and arguly the most important step, of the imaging chain is the visual inspection of the collected image. In many applications of imaging, the goal of the imaging system is a faithful reproduction of the object that has been imaged. For example, in photography, the resultant photographic print should look like the originally captured object. The imaging system may also modify the original appearance in order to enhance the reproduced image for aesthetic purposes. For this type of imaging system, it is desirle to have detectors that match the sensitivity of the human eye. However in many applications, the role of the imaging system is to transduce what is invisible or unavaille to the eye to a visual representation that can be interpreted by the human visual system. In these types of systems, important decisions must be made in how to represent the collected data in a form that is interpretle by the eye and brain of the end user. This type of information is necessarily multidimensional in nature. At least two dimensions are needed for spatial aspects of the collected data and additional dimensions are added to this framework. Color, defined as that aspect of visual perception that allows an observer to distinguish differences structure-free fields, is a natural tool for displaying this multidimensional information. In order to use color appropriately, we must understand how the visual system interprets images. The goal of this research is to discover rules and techniques for the display of multidimensional graphical information. In order to achieve this goal we need to consider the psychophysics of color and spatial perception. This knowledge must then be applied using appropriate design techniques in order to produce images that allow the end user to easily and accurately extract the relevant information from an image. The theoretical implications involved in the design of color scales has been considered in a number of articles (see for example refs. -3). However, in most cases the evaluation of these different schemes has been based on subjective judgments of effectiveness rather than by psychophysical evaluation. Although interactive exploration will most likely be necessary for implementing pseudocolor effectively, empirical results will be required in order to help guide the development of these user interfaces. This paper a presents some experimental results demonstrating how the ility of interpreting images can be affected by the encoding schemes used to generate the image. Experimental An experiment was performed in order to test how different encoding schemes affect performance in judging points on image maps. The images consisted of two digital elevation maps (DEMs) from the United States Geological Survey. The elevations were encoded using five different scales using color and intensity to represent elevation. The hypothesis was that an encoding scheme that mapped elevation in a perceptually linear manner would improve performance. In addition, the availility of the encoding

2 scheme, by way of a legend, should also affect performance differentially with the different scales. Eleven subjects participated in the experiment. b* a* Figure. The Red and Spectral color scales shown inside the outline of the CRT gamut in CIELAB space. 36 fig. and fig. 2), and 5) a gray scale based on luminance (Luminance). Figure 3 shows each of the scales. Figure 4 shows the DEM for Colorado encoded by the Spectral scale. Each map was presented with a white border and a neutral 2% surround (D65). Procedure There were two tasks. In the first task (Evaluation) a point on the image was indicated by a cross-hair (which was open in the center so as not to obscure the indicated point). The subjectsõ task was to type in the value of the indicated point. In the second task (Production) the subjects were presented with a numerical value. The subjectsõ task was to click on a point corresponding to that value in the image using a mouse. A custom cursor was used that did not obscure the chosen point. The elevations were scaled from (the lowest elevation) to (the highest elevation) which were always encoded as black and white, respectively. If the subject evaluated a point on the image as being /3 the maximum value, they would indicate by typing 33. If the subject was to click on a point with a value of 75% they were to choose a point in the image that appeared to be 3/4 of the maximum value. The subjects were not told that these images were DEMs. h* for "Spectral " scale Data Value Figure 2. Hue variation in Spectral color csale. Stimuli The DEMs were presented on a calibrated CRT using the Mapping Toolbox in MATLAB. One DEM was of an area near Rochester, NY, and the other was a more mountainous area near Denver, CO. The gamma of the monitor was approximately.6. The elevations in the two DEMs were encoded using five different schemes: ) a gray scale using digital RGB (RGB), 2) a gray scale based on with the monitor white corresponding to an = (), 3) an scale with a constant hue component ((Red ), see fig. ), 4) an scale with changing hue ((Spectral, see Figure 3. The five scales used to encode the DEMs. Each task was performed with and without an accompanying legend that indicated how the values were encoded in the image. When a legend was present it appeared as one of the bars in figure 3 with leled percentages in steps of 2%. In the no-legend conditions, only a white and black patch was displayed showing the

3 maximum and minimum values. The image maps were presented in random orientations in order to minimize memorization. The trials were presented in random order. For each of the 2 maps, 5 scales, 2 presentation conditions, and 2 tasks, there were three trials presented in a factorial design. Therefore each subject completed a total of 2 trials in out one hour. choice of encoding scheme and map location as well as some three way interactions and a 4-way interaction. Comparisons between conditions were made controlling for error due to multiple comparisons. The results are shown in Tle I. A summary of the results is plotted in figure 5. The error is plotted in reverse order on the ordinate so that higher points on the graph indicate better performance. Nonoverlapping error bars indicate significant differences. Tle I. ANOVA of error. Asterisk indicates significance at.5/5 level. Source DF P-value B (task, Evaluation or Production).2 C (location).5387 D (presence/sence of legend) * E (encoding scheme) 4 * B*C.838 B*D B*E 4.28 C*D.2327 C*E 4.4* D*E 4 * B*C*D * B*C*E 4.33* B*D*E 4.55 C*D*E 4.83 B*C*D*E 4 * Error 28 Figure 4. DEM of Colorado encoded by Spectral. Results and Discussion The performance is measured by the solute value of the difference between the target value and the chosen value as shown in equations and 2. Evaluation : Production : true value - response value error = true value target value - selected value error = target value An ANOVA was performed on the error scores using task, map location, presence/sence of legend, and encoding scheme as the four factors for analysis. The analysis revealed significant effects of choice of encoding scheme, presence or sence of a legend and an interaction between these two factors. There was also an interaction between () (2) The availility of a legend improves performance. Although this result may seem trivial, the implications are quite important. Tufte 4, for example, lists as one of his principles of graphical excellence that complex ideas should be communicated with clarity, precision and efficiency. The reliance on a legend may indicate a lack of efficiency. That is, the process of interpretation of the graphic may be verbal rather than visual. The viewer may be scanning back and forth between the legend and the map using verbal decoders and mental phrases to match colors to values. What we would like is an intuitive, visual interpretation in which relative map values can be gleaned without recourse to a legend. With no legend, Luminance encoding is far worse than the other methods. Luminance encoding would correspond to a monitor gamma of.. An RGB gray scale on a monitor with a gamma of approximately.6 improves performance over the Luminance gray scale without a legend. A monitor gamma of 2.2 corresponds closely to the encoding. Performance in judging image values is therefore dependent upon the systemõs gamma curve when using gray scales based on RGB.

4 By reanalyzing the data from the no-legend condition and the three different gray scales, we can find the best fitting exponential which matches the judgements of the evaluated or prompted values in the map. This assumes the subjects were not basing their judgments on their memory of any of the five encoding schemes. This analysis produced a best fitting curve with an exponent of.78. Further research is necessary in order to estlish the shape of the optimum lightness function and how much improvement this function provides over other curves. The addition of the hue information allows more accurate interpretation of the image values. Ware 5 found similar results. Further research is needed to determine how much additional hue information is needed to increase performance. There are many ways in which hue information could be used to supplement the gray scale encoding. This experiment only tests one possible scheme. Error RGB Legend No Legend Red Spectral Luminance Performance Figure 5. Errors for the five encoding schemes with legend present or sent. Errors are averaged over the two tasks. It is assumed that perceptual linearity in a particular color scheme will lead to better performance than color schemes based on device RGB or nonuniform color spaces. For this reason CIELAB was used in this experiment as a first approximation to a perceptually uniform space. However, an artifact of the design of the Red and Spectral used in this study is that the step sizes are not uniform, in CIELAB units, throughout the extent of the scale. This is also true of the RGB scale and the Luminance scale. In figures 6-9, the color difference in DE* and DE* 94 are shown for the RGB, Red, Spectral, and Luminance scales. Because the contribution of D is the same in both D E* and DE* 94 the curves overlap for the RGB and Luminance schemes RGB DE* 94 and DE* Figure 6. Color differences along RGB scale. Red DE* Red DE* Figure 7. Color differences along Red scale. 94

5 Spectral DE*.2 Spectral DE* Figure 8. Color differences along Spectral scale Luminance DE* 3 and DE* Figure 9. Color differences along Luminance scale. maps used in this experiment are reasonle examples of those produced by imaging systems. With a modest amount of effort, proper presentation can improve performance by well over 3%. This experiment is only a first step in understanding how to optimize the presentation of images. In this experiment only one additional dimension (elevation) is presented. More research is needed to discover methods to display multiple dimensions simultaneously. By applying our knowledge of visual processing to image display we can improve the ility of imaging to convey information clearly and accurately. References. Philip K. Robertson and John F. OÕCallaghan, The Generation of Color Sequences for Univariate and Bivariate Mapping. IEEE Computer Graphics and Applications, 6, pp , (986). 2. Haim Levkowitz. Color Theory and Modeling for Computer Graphics, Visualization, and Multimedia Application. Kluwer:Boston, B. Rogowitz and L. Treinish, Data Visualization: The End of the Rainbow. IEEE Spectrum, 35, 2, pp , (998). 4. Edward R. Tufte. The Visual Display of Quantitative Information. Graphics Press: CT, Colin Ware, Color Sequences for Univariate Maps: Theory, Experiments, and Principles. IEEE Computer Graphics and Applications, 8, pp. 4-49, (988). Biography Ethan D. Montag received his M.S. degree in Psychology from the University of Pennsylvania in 985 and a Ph.D. in Experimental Psychology from the University of California, San Diego in 99. In 994 he started a two year postdoctoral position at the Munsell Color Science Loratory at the Chester F. Carlson Center for Imaging Science at the Rochester Institute of Technology where he worked on color gamut mapping. In 996 Dr. Montag was appointed Research Assistant Professor at the Center where he pursues work in color science in the Munsell Color Science Loratory. His current interests include color gamut mapping, color vision, color tolerance measurement and the use of color in information display. Conclusion The current experiment demonstrates that the interpretation of multidimensional graphical information can be significantly affected by the way it is presented. The data